LocalRank: A Multifaceted Local Ranking Framework

Updated 8 July 2026

LocalRank is a collection of locality-constrained ranking techniques that vary by context, including client-specific LoRA rank, network centrality, local PageRank, and summarization authority scores.
It adapts methodologies such as MCDA with TOPSIS, power-iteration, and rank-1 approximations to process localized data for tailored performance across diverse applications.
Its effectiveness is demonstrated through reduced communication rounds in distributed learning, improved epidemic influence detection, faster convergence in BrowseGraph, and enhanced summarization consensus.

LocalRank is a context-dependent term used for several non-equivalent ranking constructs. In the literature represented here, it denotes a client-specific LoRA rank in distributed learning, a centrality measure in complex networks, a PageRank computed on a local BrowseGraph, and a document-cluster-specific authority score for candidate summarizers. Closely related work also studies local-only approximations of global influence from row-sums and column-sums without explicitly naming the method LocalRank (Chen et al., 2024, Bi et al., 13 Aug 2025, Trevisiol et al., 2015, Mehta et al., 2018, Bartolucci et al., 2020).

1. Scope of the term

The available literature indicates that LocalRank is not a single canonical algorithm. Rather, it is a family of locality-constrained ranking ideas whose precise mathematical object depends on the application domain.

Context	Object called LocalRank	Local information used
LoRA-enabled distributed learning	Client local rank $R_i$	Client data complexity metrics
Complex networks	Node centrality $C_i^{\rm LocalRank}$	Neighborhoods up to a neighbor’s neighbor
BrowseGraph analysis	Local PageRank $\pi^{(H)}$	A local subgraph $H\subseteq G$
Summarization aggregation	System authority score $L(k)$	Similarity among candidate summaries
Partial-information influence ranking	Rank-1 influence approximation $\hat I_i$	Row-sums $r_i$ , column-sums $c_i$

Across these usages, “local” refers to restricted observability or bounded context: client-specific data statistics, local neighborhoods, local browsing traces, cluster-specific summary similarity, or aggregate degree/strength information. A plausible implication is that LocalRank is best understood as a design pattern for ranking under incomplete or localized information, rather than as a single named estimator.

2. LocalRank as client-specific LoRA rank in distributed learning

In LoRA-enabled distributed learning, each client $i$ freezes the bulk of a large pre-trained model and trains two low-rank adaptation matrices $A_i\in\mathbb R^{d\times R_i}$ and $C_i^{\rm LocalRank}$ 0. The integer $C_i^{\rm LocalRank}$ 1 is the client’s local rank. The dimension of the rank- $C_i^{\rm LocalRank}$ 2 LoRA update is proportional to $C_i^{\rm LocalRank}$ 3, rather than $C_i^{\rm LocalRank}$ 4. A larger $C_i^{\rm LocalRank}$ 5 increases model capacity, but also raises local computation, memory footprint, and risk of over-fitting; a smaller $C_i^{\rm LocalRank}$ 6 limits cost but may under-fit complex data and slow global convergence. In practice, the AutoRank framework lets each client scale a global rank $C_i^{\rm LocalRank}$ 7 via

$C_i^{\rm LocalRank}$ 8

where $C_i^{\rm LocalRank}$ 9 is a normalized rank-scale-ratio computed by the server (Chen et al., 2024).

The motivating theory is the bias–variance trade-off. For client $\pi^{(H)}$ 0, the local squared-loss generalization error is written as

$\pi^{(H)}$ 1

with model complexity $\pi^{(H)}$ 2 and data complexity $\pi^{(H)}$ 3. Under mild assumptions on the shape of bias and variance curves, the analysis yields $\pi^{(H)}$ 4: more complex data requires a larger local rank to remain on the “sweet-spot” of the bias–variance trade-off. AutoRank operationalizes this with three proxy metrics for $\pi^{(H)}$ 5: loss-entropy $\pi^{(H)}$ 6, label-distribution entropy $\pi^{(H)}$ 7, and the Gini–Simpson index $\pi^{(H)}$ 8. Higher values of these metrics indicate more complex, heterogeneous, or larger-scale data.

Rank assignment is then cast as MCDA using TOPSIS. The server forms a decision matrix over the three metrics, normalizes it, weights each criterion using the CRITIC method, determines the ideal and anti-ideal points, computes the separations $\pi^{(H)}$ 9 and $H\subseteq G$ 0, and assigns each client a relative closeness

$H\subseteq G$ 1

Larger $H\subseteq G$ 2 means that the client’s data profile is closer to the “worst” (most complex) ideal and therefore receives a higher rank. The server min–max normalizes the $H\subseteq G$ 3 values to $H\subseteq G$ 4 and broadcasts the resulting $H\subseteq G$ 5.

The reported experiments concern highly heterogeneous, double-imbalanced non-IID scenarios on MNIST, FMNIST, CIFAR-10, and CINIC-10. AutoRank yields 10–20 % fewer communication rounds to reach peak accuracy than fixed-rank FedLoRA or manually tuned heterogeneous-rank RBLA; global accuracy improves by 0.3–2 % in MLP tasks and up to 3 % in CNN tasks; and total trainable parameters are reduced by 15–20 % without loss of accuracy. Alternative metric sets for unlabeled data still outperform homogeneous rank and match or exceed manual schemes. In this usage, LocalRank is a model-capacity control variable rather than a node-ranking score.

3. LocalRank as a centrality measure in complex networks

In the network-science literature summarized here, LocalRank is a node centrality defined by

$H\subseteq G$ 6

where $H\subseteq G$ 7 is the set of 1st-order neighbors of node $H\subseteq G$ 8, and

$H\subseteq G$ 9

is the total number of $L(k)$ 0’s 1st- and 2nd-order neighbors. The procedure is local in the sense that it expands only through neighborhoods, but it goes beyond degree by aggregating the reach of neighbors of neighbors. On an undirected, unweighted sparse network, computing $L(k)$ 1 for all $L(k)$ 2 takes $L(k)$ 3, and the overall complexity is $L(k)$ 4 (Bi et al., 13 Aug 2025).

Empirically, this LocalRank behaves as a meso-scale centrality. Over 80 real-world networks, it clusters with Subgraph Centrality, Eigenvector Centrality, and Closeness in “Group 2.” Within that group, all pairwise Kendall’s $L(k)$ 5 exceed $L(k)$ 6; specifically,

$L(k)$ 7

Its correlations with Degree, Katz, IC, LDD, MDD, H-index, and Coreness lie roughly in $L(k)$ 8, while correlations with Collective Influence, Cycle Ratio, Betweenness, Leverage Centrality, and Eccentricity are typically $L(k)$ 9.

The same study evaluates LocalRank under the SIR epidemic model with recovery probability $\hat I_i$ 0 and infection rate $\hat I_i$ 1. For the single-seed task, where influence $\hat I_i$ 2 is the average final outbreak size from node $\hat I_i$ 3 alone, LocalRank attains the highest Precision curve for all $\hat I_i$ 4, and its median Kendall’s $\hat I_i$ 5 is approximately $\hat I_i$ 6, above Subgraph Centrality and Katz. For the set-influence task, where the top- $\hat I_i$ 7 nodes by centrality are seeded simultaneously, LocalRank drops near the bottom for larger $\hat I_i$ 8. The reason given is spatial clustering: on the benchmark rt_tlot network, the top 1% nodes by LocalRank cluster in the core, and the average pairwise distance

$\hat I_i$ 9

is among the smallest of the 16 measures. This directly illustrates a common misconception: a centrality that is strong for identifying the single most influential node need not be strong for selecting influential node sets.

4. LocalRank on the BrowseGraph

In BrowseGraph analysis, LocalRank denotes PageRank computed on a local subgraph rather than on the global graph. Let $r_i$ 0 be the directed, weighted global BrowseGraph and $r_i$ 1 a local subgraph induced by a subset of sessions. If $r_i$ 2 is the local transition matrix and $r_i$ 3 the teleport distribution restricted to $r_i$ 4, the LocalRank $r_i$ 5 is defined by

$r_i$ 6

It can be computed by the standard PageRank power iteration on $r_i$ 7, and in closed form satisfies

$r_i$ 8

(Trevisiol et al., 2015).

The central question is not merely how to compute $r_i$ 9, but how far it diverges from the restriction of the global PageRank $c_i$ 0 to $c_i$ 1. This divergence is measured by Kendall’s $c_i$ 2. To predict $c_i$ 3 from local information alone, the study computes 62 structural features of $c_i$ 4, including size and connectivity, degree statistics, weighted-degree statistics, clustering and centralization, LocalRank distribution statistics, and assortativity. A Random Forest regression model achieves the best MSE, approximately $c_i$ 5, and weighted-degree statistics alone predict $c_i$ 6 almost as well as all features together. On held-out graphs, the predicted ordering of $c_i$ 7 over the seven referrer-based subgraphs reaches Spearman $c_i$ 8; with all features, $c_i$ 9.

The empirical setting is Yahoo! News server logs from 2013, with 3.8M pageviews and 1.88B transitions. Seven local subgraphs are built from sessions whose first referrer is Google, Yahoo, Bing, Facebook, Twitter, Reddit, or Homepage. Under breadth-first “growing rings,” where one iteratively expands the local graph by adding all out-neighbors, Kendall’s $i$ 0 reaches $i$ 1– $i$ 2 after 2–3 rings. If only the top 5–30% of new nodes by local PageRank are added at each ring, convergence is faster initially, with $i$ 3 versus $i$ 4, although it plateaus slightly below $i$ 5 in the long run. In this usage, LocalRank is not a distinct centrality family but the localized version of PageRank under restricted graph access.

5. LocalRank in aggregation of summarization systems

In extractive summarization, LocalRank is a document-cluster-specific authority score assigned to each candidate summarizer. Let $i$ 6 be the summary produced by candidate system $i$ 7 for a given cluster, and let $i$ 8 denote the cosine similarity between bag-of-words representations of summaries. One builds a fully connected, weighted graph over candidate systems, with edge weight $i$ 9. The LocalRank scores $A_i\in\mathbb R^{d\times R_i}$ 0 satisfy

$A_i\in\mathbb R^{d\times R_i}$ 1

If $A_i\in\mathbb R^{d\times R_i}$ 2 is the row-normalized similarity matrix with zero diagonal, then $A_i\in\mathbb R^{d\times R_i}$ 3. Computation proceeds by summary generation, pairwise cosine similarity, graph construction, uniform initialization, and PageRank-style iteration until convergence (Mehta et al., 2018).

The score is then used to weight sentence selection across candidate summaries. For sentence $A_i\in\mathbb R^{d\times R_i}$ 4 in summary $A_i\in\mathbb R^{d\times R_i}$ 5,

$A_i\in\mathbb R^{d\times R_i}$ 6

Candidate sentences are sorted by $A_i\in\mathbb R^{d\times R_i}$ 7, and selection continues until the 100-word budget is reached; any sentence whose cosine similarity exceeds $A_i\in\mathbb R^{d\times R_i}$ 8 with an already selected sentence is discarded for redundancy control.

This LocalRank is explicitly contrasted with GlobalRank, a corpus-wide reputation score $A_i\in\mathbb R^{d\times R_i}$ 9 estimated on DUC2002 from average ROUGE-1 recall, and with HybridRank,

$C_i^{\rm LocalRank}$ 00

with $C_i^{\rm LocalRank}$ 01. On DUC2003, the reported ROUGE-1 / ROUGE-2 / ROUGE-4 recalls are $C_i^{\rm LocalRank}$ 02 for LocalRank and $C_i^{\rm LocalRank}$ 03 for HybridRank, with the HybridRank gains marked as statistically significant by a two-sided sign test with $C_i^{\rm LocalRank}$ 04. The data therefore support a specific interpretation: cluster-specific local agreement among systems is useful, but combining it with corpus-level reputation is stronger than using LocalRank alone.

Bartolucci et al. do not introduce a method explicitly named LocalRank. They instead study when global “influence” measures can be reconstructed from local statistics, which makes the work directly relevant to local ranking methods. Let $C_i^{\rm LocalRank}$ 05 be a non-negative, sub-stochastic interaction matrix and define the exact influence vector

$C_i^{\rm LocalRank}$ 06

With row-sums $C_i^{\rm LocalRank}$ 07 and column-sums $C_i^{\rm LocalRank}$ 08, the matrix $C_i^{\rm LocalRank}$ 09 is replaced by the unique maximum-entropy rank-1 proxy

$C_i^{\rm LocalRank}$ 10

Applying the Sherman–Morrison formula gives a closed form for the double-constraint approximation,

$C_i^{\rm LocalRank}$ 11

and, when only row-sums are known, the single-constraint approximation

$C_i^{\rm LocalRank}$ 12

(Bartolucci et al., 2020).

The complexity is $C_i^{\rm LocalRank}$ 13 time and $C_i^{\rm LocalRank}$ 14 memory to read $C_i^{\rm LocalRank}$ 15 once, or $C_i^{\rm LocalRank}$ 16 memory if only the sums are known. No fully explicit worst-case bound is given, but the approximation is controlled by the spectral gap: when the Perron–Frobenius eigenvalue $C_i^{\rm LocalRank}$ 17 dominates $C_i^{\rm LocalRank}$ 18, one observes

$C_i^{\rm LocalRank}$ 19

Empirical tests on synthetic ER and scale-free graphs, and on real networks including Facebook ego, arXiv GR-QC collaboration, OpenFlights, the Stanford web corpus, WIOD input–output tables, and the St. Marks food web, show Pearson and Spearman correlations exceeding $C_i^{\rm LocalRank}$ 20 and often $C_i^{\rm LocalRank}$ 21, except trophic levels at approximately $C_i^{\rm LocalRank}$ 22. The doubly constrained form always outperforms the single-constraint form, while performance deteriorates on very sparse networks when the spectral gap becomes small.

7. Comparative interpretation and limitations

A recurrent misconception is that LocalRank denotes one standard ranking algorithm. The literature summarized here shows instead that the term spans at least four distinct mathematical objects and one closely related local-only approximation paradigm. The shared idea is locality under constraint, not a fixed formula.

The role of locality also differs sharply by domain. In AutoRank, locality is client-specific model capacity allocation under double-imbalanced non-IID data. In complex networks, locality means bounded-hop structural information, and the key trade-off is between identifying a single influential node and identifying dispersed influential node sets. In BrowseGraph analysis, locality means ranking on an incomplete subgraph and estimating divergence from global PageRank. In summarization, locality means cluster-specific consensus among candidate systems, which is improved further by combining it with a global reputation signal. In partial-information influence ranking, locality reduces the problem to row-sums, column-sums, and a rank-1 proxy.

These differences imply that results do not transfer automatically across contexts. A high-performing LocalRank for epidemic spreading is not thereby a good model-selection rule for federated learning, and a local PageRank on browsing traces is not equivalent to a neighborhood-based network centrality. The strongest domain-specific limitations reported here are similarly heterogeneous: LocalRank centrality clusters its top-ranked nodes too tightly for seed-set selection; local BrowseGraph ranking can diverge from the global ranking unless expansion or predictive calibration is used; and rank-1 local approximations degrade when networks are exceedingly sparse. The term is therefore best treated as a contextual label attached to different local ranking mechanisms rather than as a unified theory.